Copyright, ownership, and everything that is unresolved

Three questions that keep getting merged into one

Almost every confused conversation about AI and copyright is really three conversations happening at once. Separate them and each becomes tractable — and it becomes clear how much is genuinely unresolved.

1. Was it lawful to train the model on copyrighted material?
2. Does a particular output infringe somebody's copyright?
3. Does anyone own the output, and can it be protected?

Nothing in this lesson is legal advice, and the position differs sharply by country. What follows is the shape of the disagreement, so that you can recognise which question you are in and know when to ask somebody qualified.

Question one: training

This is being litigated and legislated simultaneously, and the answers diverge by jurisdiction.

The European Union has text-and-data-mining exceptions in the 2019 Digital Single Market Directive: broad for research bodies, and for other uses subject to rightsholders reserving their rights — an opt-out mechanism whose practical effect is contested.

Japan has a comparatively permissive provision allowing use of works for machine analysis in many circumstances.

The United Kingdom has a narrow exception limited to non-commercial research. A broader one was proposed and dropped after strong opposition from creative industries, and the question has been through further consultation without settling.

The United States has no specific exception; the argument runs through fair use, and multiple cases are live. Different courts have reached different conclusions on different facts, and appeals are outstanding.

For you at work, question one is mostly somebody else's problem — with one exception. If a court somewhere decides against a provider, the tool you depend on may change or become unavailable, and a process built entirely around one vendor is exposed to that. That is a continuity risk, not a copyright risk.

Question two: does the output infringe

Separate, and more immediately practical. A model can produce material substantially similar to something in its training data. The risk is highest where the training material is distinctive and the request points at it — images in a named living artist's style, a well-known character, a distinctive melody, a block of code that exists verbatim in a public repository under a licence with conditions.

For ordinary business prose it is low but not zero. For images, music and code it is real enough to matter.

Practical consequences, none of them controversial:

• Do not ask for work "in the style of" a named living creator for anything commercial.
• Treat generated code as code of unknown provenance: check anything that looks familiar, and be aware that some open licences impose obligations that survive copying.
• Trademark and passing off are untouched by any of this. A generated logo that resembles an existing mark is a trademark problem regardless of copyright.

Some providers offer contractual indemnities to business customers covering copyright claims arising from outputs. They are real, and they are conditional — typically on using the provider's own filters and not deliberately steering towards infringement. If your organisation is relying on one, somebody should have read the conditions.

Question three: who owns it

This is the one that surprises people, and the divergence is stark.

The United States Copyright Office has taken the position that copyright protects works of human authorship, and that purely machine-generated material is not registrable; works combining human authorship with generated elements may be registered as to the human contribution, which must be disclosed. Litigation over an application naming a machine as author failed on the same principle.

The United Kingdom has an unusual provision — section 9(3) of the Copyright, Designs and Patents Act 1988 — treating the author of a computer-generated work as the person who made the arrangements necessary for its creation, with a shorter term. Whether it should be retained has been the subject of consultation.

China has produced court decisions finding sufficient human input in AI-assisted images to attract protection. India and other jurisdictions are working through their own positions.

So the honest summary is: whether your generated output is protectable at all depends on where you are, how much you contributed, and how the current round of cases and consultations settles. Nobody should tell you this is resolved.

What to do at work while it is unresolved

• Read your client contracts. Many now contain clauses about AI use, and warranties that deliverables are original and owned. You may be warranting something you cannot verify.
• Do not build a protectable asset on generated material alone. A brand mark you may be unable to register is a poor foundation, and the register is what you enforce with.
• Keep the human contribution evident and recorded. The record-keeping habits from the previous block — what was generated, from what, edited by whom — are exactly what a registration or a dispute would ask for.
• Know that ownership and confidentiality are different duties. Material you own outright can still be material you are contractually forbidden to disclose, and pasting it into a tool is disclosure.
• When it matters commercially, ask a qualified adviser in your jurisdiction. Not the model, which will answer confidently and in the register of whichever country's material dominated its training.